Correction on my score conversions: calibrating against a new run's displayed score (loop 1: 44.35 fps shown as 4434), the factor is fps×100, not ×96.6 as I used. Revised: fans-auto ~4122 first / ~3583 sustained (same regime as BETA15's 3983/3432, ~4% higher on the larger chassis — not digit-for-digit); fans-100% sustained ~4121, i.e. ~6% below the 4392 launch figure rather than ~9%, with ~27°C ambient here likely explaining part of it. Qualitative picture unchanged.

Correction on my score conversions: calibrating against a new run's displayed score (loop 1: 44.35 fps shown as 4434), the factor is fps×100, not ×96.6 as I used. Revised: fans-auto ~4122 first / ~3583 sustained (same regime as BETA15's 3983/3432, ~4% higher on the larger chassis — not digit-for-digit); fans-100% sustained ~4121, i.e. ~6% below the 4392 launch figure rather than ~9%, with ~27°C ambient here likely explaining part of it. Qualitative picture unchanged.

My pleasure, and my necessity 😅! I've been working for a while on developing my version of Metal Flash Attention for MLX, that includes many functions not present in MLX, and while testing and benching my last release, I noticed big performance drops compared to the previous release. I couldn't believe that the entire performance drop was due to macOS 27 beta (I upgraded to test Metal 4.1) and had to investigate whether it was a "me" problem, or something else. turns out it's something else!